6.2 KiB
| slug | sidebar_position | sidebar_label | title |
|---|---|---|---|
| /en/operations/caches | 65 | Query Result Cache [experimental] | Query Result Cache [experimental] |
Query Result Cache [experimental]
The query result cache allows to compute SELECT queries just once and to serve further executions of the same query immediately from the cache. Depending on the type of the queries, this can dramatically reduce latency and resource consumption of the ClickHouse server.
Background, Design and Limitations
Query result caches are generally either transactionally consistent or inconsistent.
- In transactionally consistent caches, the database invalidates/discards cached query results if the result of the SELECT query changes or potentially changes. In ClickHouse, operations which change the data include inserts/updates/deletes in/of/from tables or collapsing merges. Transactionally consistent caching is especially suitable for OLTP databases, for example MySQL (which removed query result cache after v8.0) and Oracle.
- In transactionally inconsistent caches, slight inaccuracies in query results are accepted under the assumption that all cache entries are assigned a validity period after which they expire (e.g. 1 minute) and that the underlying data changes only little during this period. This approach is overall more suitable for OLAP databases. As an example where transactionally inconsistent caching is sufficient, consider an hourly sales report in a reporting tool which is simultaneously accessed by multiple users. Sales data changes typically slowly enough that the database only needs to compute the first report (represented by a SELECT query). Further queries can be served directly from the query result cache. In this example, a reasonable validity period could be 30 min.
Transactionally inconsistent caching is traditionally provided by client tools or proxy packages interacting with the database. As a result, the same caching logic and configuration is often duplicated. With ClickHouse's query result cache, the caching logic moves to the server side. This reduces maintenance effort and avoids redundancy.
:::warning This is an experimental feature that should not be used in production. There are known cases (e.g. in distributed query processing) where wrong results are cached/stored. :::
Usage Examples and Configuration Settings
The query/user/profile-level parameter enable_experimental_query_result_cache controls whether query results are inserted or retrieved from the cache. For example, the first execution of query
SELECT expensive_calculation(A, B, C)
FROM T
SETTINGS enable_experimental_query_result_cache = true;
will store the query result into the query result cache and subsequent executions will retrieve the result directly from the cache.
It is sometimes desirable to use the query result cache only passively, i.e. to read from it but not write in it. Parameter enable_experimental_query_result_cache_passive_usage instead of 'enable_experimental_query_result_cache' can be used for that.
For maximum control, it is generally recommended to enable caching on a per-query basis. It is also possible to activate caching at user/profile level but users should keep in mind that all SELECT queries may then return outdated results.
To clear the query result cache, use statement SYSTEM DROP QUERY RESULT CACHE. The content of the query result cache is displayed in
system table SYSTEM.QUERY_RESULT_CACHE. The number of query result cache hits and misses are shown as events "QueryResultCacheHits" and
"QueryResultCacheMisses" in system table SYSTEM.EVENTS. Both counters are updated only for SELECT queries which run with settings
"enable_experimental_query_result_cache = true" or "enable_experimental_query_result_cache_passive_usage = true". In particular, all other
queries do not increment the cache miss counter.
The cache exists once per ClickHouse server process but cache results are by default not shared between users (see below).
Query results are referenced in the cache by the Abstract Syntax Tree (AST) of their
query. This means that caching is agnostic to upper/lowercase, for example SELECT 1 and select 1 are treated as the same query.
If the query was aborted by an exception or cancelled, no entry is written into the query result cache.
Further Configuration Options:
Different server configuration options are available to specify the size of the query result cache, the maximum number of cache entries and the maximum size of a cache entry in bytes and in records.
To define how long a query must run at least such that its result is cached, use setting query_result_cache_min_query_duration.
To control how often a query needs to run until its result is cached, use setting query_result_cache_min_query_runs.
To specify the validity period after which cache entries become stale, use setting query_result_cache_ttl.
Results of queries with non-deterministic functions such as rand() and now() are not cached by default. This behavior can be overruled using setting query_result_cache_store_results_of_queries_with_nondeterministic_functions.
Query cache entries are not shared between users due to security reasons. For example, user A must not be able to bypass a row policy on a table by running the same query as another user B for whom no such policy exists. If nevertheless necessary, cache entries can be marked accessible by other users (i.e. shared) using setting [query_result_cache_share_between_users]{settings/settings.md#query-result-cache-share-between-users}.